Files
MeltMind/docs/claude/stage-research-plan-2026-11-18.md
T

5.3 KiB
Raw Blame History

Stage research plan (Claude, 2026-11-18)

Yanqi's direction (via TRON, 2026-11-18): keep the current plan; aim first for SOTA on 3D benchmarks (industrial use is a bonus); research graph-network history and frontier, pretraining, similar-physics data, sensible later use of RL; investigate procedural generation of many molding simulations; use GINO and Greif data first and try to beat their results. Comparisons must be same task, same data, fair split. Data generation / pretraining / RL are directions to explore, not validated or authorised expensive jobs.

Undermind deep searches (workspace "Injection Molding Hele Shell Manifold Neural Solvers", folder /phase3 research plan 2026-11-18/): (1) programmatic generation of large molding simulation datasets; (2) graph / neural-operator surrogates on 3D meshes, history, frontier and pretraining; (3) RL with physics simulators for manufacturing and data acquisition.

Tracks

Track Question First step (owner) Gate before anything expensive
A. Greif SOTA Does H2 beat Gre24 on Gre24's own protocol? Our grouped protocol is stricter, so the paper's numbers (Table C.6: 3.38 % median RRMSE at 100 trees / depth 8; 3.31 % at depth 6, the lowest in that table; MeltAssist 2026-11-18) are not comparable with ours. The exact split seed is not published, so Track A compares A and H2 on the same re-created split(s), never against the paper number alone Phase 3 continues unchanged (item 1 running). MeltAssist: extract Gre24's exact split, metric and reported numbers. MeltAstra: then evaluate A and H2 on that protocol Reproduction of Gre24's own number on its protocol first (we have 3.96 % on v4)
B. GINO benchmark Does our mesh model match or beat GINO and later SOTA on GINO's public 3D tasks (ShapeNet car pressure 500/111; Ahmed body 500/51)? MeltAssist: verify data sources, splits, metrics and the current reported SOTA table. MeltAstra: download, reproduce the GINO baseline Reproduce a published baseline within its reported error before claiming anything
C. Data generation Can we generate thousands of filling simulations cheaply enough and faithfully enough? Deep search 1 + MoldAstra: a fidelity ladder (eikonal proxy → Hele-Shaw / 2.5D → open-source 3D VOF → commercial) with cost per part. Question to Yanqi: is a Moldflow or Moldex3D licence available? Greif's own data came from scripted Moldflow 2023, which is the fastest faithful route A pilot on about 10 Greif parts comparing the generator with the Moldflow labels, with a pre-registered fidelity threshold
D. Pretraining Does pretraining (synthetic labels from C, or self-supervised geometry) improve H2 on Greif at equal fine-tune budget? MoldAstra: evaluation design (equal compute, same splits). Self-supervised geometry pretraining needs no generator and can be tested first Only after phase-3 item 1 shows a data limit, or after C passes its fidelity gate
E. RL Where does RL add value over Bayesian optimisation / active learning? Deep search 3 only. Candidate later uses: choosing which simulations to run (active learning), gate placement as a downstream application No RL job until a generator exists

Order

  1. Now: phase 3 continues (item 1 → seed 1 → c → b → ensemble). Literature (deep searches, MeltAssist synthesis), Track A and B protocol extraction, and Track C/D specs proceed on paper in parallel.
  2. Next compute after item 1: Track A on Gre24's protocol (cheap), then Track B GINO reproduction.
  3. Tracks C/D pilots only after their gates and the licence answer.

Update after MeltAssist synthesis (2026-11-18, parts 1-3)

  • Track A protocol frozen: five independent random 314/315 splits (our published seeds; protocol-level re-creation), per-part RMSE / max(full truth) proxy, min/median/max over test parts averaged over splits; A and H2 paired on the same splits; family leakage of random splits reported.
  • Track B frozen: original GINO decoder protocol only (ShapeNet 611, 500/111; Ahmed 551, 500/51; de-normalised relative L2). Targets: GINO 7.12 % / 8.31 % and same-protocol results confirmed by MeltAssist; other splits in a context table. ShapeNet data (Zenodo 13936501) pre-authorised after the ID-mapping check; Ahmed on hold until the original 551 source is confirmed. Reproduce GINO with the paper recipe first.
  • Track C evidence: batch Moldflow automation is proven (Greif; Per23: 1,080 runs, about 22 min/run, 3 geometries). openInjMoldSim's high-fidelity example cost 34 h on 192 cores, so open-source 3D VOF is not a cheap mass generator. Report independent geometries and conditions separately, and acceptance rate and total cost per valid label.
  • Track D evidence: GeoPT (13,463 ShapeNet geometries × random dynamics fields, no CFD labels; 144-576 GPU-h) reports 20-60 % label savings on several 3D tasks (not molding). Geometry-only pretraining is therefore the first D arm, compared with scratch at equal fine-tune budget, pretraining cost reported separately.
  • Track E: RL evidence (Kim25 surrogate-environment process optimisation; Sza22 VARTM layup) is narrow and not thermoplastic filling; stays literature-only until a trusted environment exists.
  • Idle GPUs: approved queue items may use GPUs that item 1 does not need, each after its own review: Track A H2 first, then seed 1.